"Artificial evolution produces solutions that programmers couldn't have even dreamed of"
I've been thinking about this lately. I don't know how BigDog was conceived and produced, but it seems plausible that one could have used a system similar to what appears in the video below. We can get incredibly lifelike motion from simple parts by simulating, on the computer, a creature evolving in a physical environment.
I know that the same kinds of algorithms have been used to evolve neural networks. The problem is that they have been very limited in scope because the networks are used to solve specific problems. However, what if we used the most recent insights from cognitive science (about embodiment) and created a simulated creature with both a body and a complex neural network (or networks) as a control mechanism? What would happen, given some constraints informed by theory, if we evolved such a creature? We may wind up getting intelligent behavior way more quickly than if we tried to program it piece by piece.
Of course, the kinds of constraints would be enormously complex. We would have to provide specifications on perception, motor control, etc.., plus we would have to choose and code the relevant selection rule. For the examples in the video the selection rules are simple: Get as far off the ground as possible, or move as far from the original location as possible. Real creatures have evolved amidst an incredibly complex and interacting array of selective pressures. Not only this, but we would have to provide more in the simulated environment than a flat ground. What about other organisms to interact with, objects in the environment, even weather conditions?
Youtube link here